Fuzzy hash of behavioral results

Abstract

A computerized method for classifying objects in a malware system is described. The method includes detecting behaviors of an object for classification after processing of the object has begun. data associated with the detected behaviors is collected, and a fuzzy hash for the received object is generated. The generation of the fuzzy hash may include (i) removing a portion of the data associated with the detected behaviors, and (ii) performing a hash operation on a remaining portion of the data associated with the detected behaviors. Thereafter, the fuzzy hash for the received object is compared to a fuzzy hash of an object in a preexisting cluster to generate a similarity measure. The received object is associated with the preexisting cluster in response to determining that the similarity measure is above a predefined threshold value. Thereafter, the results are reported.

Description

CROSS-REFERENCE OF RELATED APPLICATIONS

This application is a continuation of U.S. patent application Ser. No. 14/042,454 filed on Sep. 30, 2013, now U.S. Pat. No. 9,294,501 issued Mar. 22, 2016. The entire contents of U.S. patent application Ser. No. 14/042,454 are incorporated by reference herein.

FIELD

Embodiments of the disclosure relate to the field of network security. More specifically, one embodiment of the disclosure relates to a system, apparatus, and method for classifying a suspect object in a malware system using a fuzzy hash of behaviors of the suspect object and clusters of previously classified objects.

GENERAL BACKGROUND

Over the last decade, malicious software (malware) has become a pervasive problem for Internet users. In some situations, malware is a program, file, or digital data object that is embedded within downloadable content and designed to adversely influence (i.e., attack) normal operations of a computer. Examples of different types of malware may include bots, computer viruses, worms, Trojan horses, spyware, adware, or any other programming that operates within the computer without permission.

For instance, content may be embedded with objects associated with a web page hosted by a malicious web site. By downloading this content, malware causing another web page to be requested from a malicious web site may be unknowingly installed on the computer. Similarly, malware may also be installed on a computer upon receipt or opening of an electronic mail (email) message. For example, an email message may contain an attachment, such as a Portable Document Format (PDF) document, with embedded executable malware. Also, malware may exist in files infected through any of a variety of attack vectors, which are uploaded from the infected computer onto a networked storage device such as a file share.

As development of malware has progressed, hackers have developed malware that share similarities with other malware objects, but maintain some dissimilarities. Accordingly, these “similar” malware objects may be in the same malware family, but traditional malware and anti-virus protection systems may fail to properly classify each object in the family as malware based on these differences. For example, traditional malware detection and classification techniques may employ a direct comparison of a suspect object with known malware objects in an attempt to reveal an exact match. However, if the suspected malware object has not been previously detected and analyzed (e.g., zero-day malware threats), these direct comparison techniques will fail to classify the object as malware even if “similar” objects have been previously classified as malware. Accordingly, traditional malware classification and analysis techniques may prove inaccurate and inefficient as these techniques do not accommodate for small difference between malware objects within a family of malware.

BRIEF DESCRIPTION OF THE DRAWINGS

Embodiments of the invention are illustrated by way of example and not by way of limitation in the figures of the accompanying drawings, in which like references indicate similar elements and in which:

FIG. 1 is an exemplary block diagram of a communication system deploying a plurality of malicious content detection (MCD) systems according to one embodiment of the invention.

FIG. 2 is an exemplary block diagram of a MCD system according to one embodiment of the invention.

FIG. 3 is a diagram of a method for classifying objects using fuzzy hashes of previously classified objects according to one embodiment of the invention.

FIG. 4A shows an example user interface for entering information for a suspect object according to one embodiment of the invention.

FIG. 4B shows the example user interface of FIG. 4A after a warning message has been returned to a user according to one embodiment of the invention.

FIG. 5 is a diagram of a set of objects assigned to a set of clusters according to one embodiment of the invention.

FIG. 6 is a diagram of a set of objects known as malware, known as non-malware, or with an unknown status and assigned to a set of clusters according to one embodiment of the invention.

FIG. 7A is a diagram of a suspect object being added to a preexisting cluster according to one embodiment of the invention.

FIG. 7B is a diagram of a suspect object being added to a new cluster according to one embodiment of the invention.

DETAILED DESCRIPTION

I. Overview

In one embodiment of the invention, a communication system is provided that includes a plurality of malicious content detection (MCD) systems communicatively coupled to a management system via a network. Each of the MCD systems may detonate, execute, open, or otherwise process a suspected malware object such that the suspect object conducts/performs a set of behaviors. These behaviors are collected and recorded such that further analysis with objects previously analyzed and assigned to clusters may be performed. In one embodiment, the recorded behavior data for the suspect object may be used to generate a fuzzy hash. A fuzzy hash allows the comparison of objects to determine similarity of the objects instead of necessarily a direct match. In comparison, traditional hashing techniques only allow a comparison of objects to determine an exact match. By allowing the determination of “similar” objects, fuzzy hashes afford a greater leniency in classification and categorization of objects that might be slightly different but otherwise share important characteristics.

As alluded to above, the fuzzy hash of the suspect object is compared against fuzzy hashes of one or more objects in one or more clusters. In one embodiment, machine learning may be utilized to determine a “similar” object in a cluster. Upon detection of a “similar” object, the suspect object may be associated with the cluster and classified based on information attached to the cluster. For example, the suspect object may be classified as malware, non-malware, or with an unknown status based on the classification of objects within the cluster. In some embodiments, the suspect object may be assigned a malware family name associated with the cluster.

As described above, fuzzy hash techniques may be used to group “similar” objects in clusters for further analysis and classification. This similarity matching provides 1) greater flexibility in analyzing potential malware objects, which may share multiple characteristics and behaviors but are also slightly different from previously classified objects, 2) a more efficient technique for classifying/assigning attributes to objects (e.g., malware family names), and 3) increase accuracy in identifying malware.

II. Terminology

In the following description, certain terminology is used to describe features of the invention. For example, in certain situations, the terms “logic” and “engine” are representative of hardware, firmware or software that is configured to perform one or more functions. As hardware, logic may include circuitry such as processing circuitry (e.g., a microprocessor, one or more processor cores, a programmable gate array, a microcontroller, an application specific integrated circuit, etc.), wireless receiver, transmitter and/or transceiver circuitry, semiconductor memory, combinatorial logic, or other types of electronic components.

As software, logic may be in the form of one or more software modules, such as executable code in the form of an executable application, an application programming interface (API), a subroutine, a function, a procedure, an applet, a servlet, a routine, source code, object code, a shared library/dynamic load library, or one or more instructions. These software modules may be stored in any type of a suitable non-transitory storage medium, or transitory storage medium (e.g., electrical, optical, acoustical or other form of propagated signals such as carrier waves, infrared signals, or digital signals). Examples of non-transitory storage medium may include, but is not limited or restricted to a programmable circuit; a semiconductor memory; non-persistent storage such as volatile memory (e.g., any type of random access memory “RAM”); persistent storage such as non-volatile memory (e.g., read-only memory “ROM”, power-backed RAM, flash memory, phase-change memory, etc.), a solid-state drive, hard disk drive, an optical disc drive, or a portable memory device. As firmware, the executable code is stored in persistent storage.

The term “content” generally refers to information transmitted over a network as one or more messages, namely a grouping of information that comprises a header and a payload, such as any of the following: a packet; a frame; a stream being a sequence of packets or frames; an Asynchronous Transfer Mode “ATM” cell; or any other series of bits having a prescribed format. An “object” may be construed as a portion of the content, namely information within one or more of the messages. The “payload” is generally defined as including the data associated with the message such as text, executable software, an image, audio, video, a Uniform Resource Locator (URL), or other types of digital data. The “header” is generally defined as a part of the message that includes control information. However, the specific types of control information depend on the content/object type.

For network traffic, such as data transmitted in accordance with a Hypertext Transfer Protocol (HTTP), HyperText Markup Language (HTML) protocol, the header may include source and destination Internet Protocol (IP) addresses (e.g., IPv4 or IPv6 addressing) and/or source and destination port information.

Another example of content or objects includes email, which may be transmitted using an email protocol such as Simple Mail Transfer Protocol (SMTP), Post Office Protocol version 3 (POP3), or Internet Message Access Protocol (IMAP4). A further example of content or objects includes an Instant Message, which may be transmitted using Session Initiation Protocol (SIP) or Extensible Messaging and Presence Protocol (XMPP) for example. Yet another example of content or objects includes one or more files that are transferred using a data transfer protocol such as File Transfer Protocol (FTP) for subsequent storage on a file share. Where the content or object is email, Instant Message or a file, the header may include the sender/recipient address, the sender/recipient phone number, or a targeted network location of the file, respectively.

The term “malware” is directed to software that produces an undesirable behavior upon execution, where the behavior is deemed to be “undesirable” based on customer-specific rules, manufacturer-based rules, or any other type of rules formulated by public opinion or a particular governmental or commercial entity. This undesired behavior may include a communication-based anomaly or an execution-based anomaly that (1) alters the functionality of an electronic device executing that application software in a malicious manner; (2) alters the functionality of an electronic device executing that application software without any malicious intent; and/or (3) provides an unwanted functionality which is generally acceptable in other context.

The term “transmission medium” is a communication path between two or more systems (e.g. any electronic devices with data processing functionality such as, for example, a security appliance, server, mainframe, computer, netbook, tablet, smart phone, router, switch, bridge or router). The communication path may include wired and/or wireless segments. Examples of wired and/or wireless segments include electrical wiring, optical fiber, cable, bus trace, or a wireless channel using infrared, radio frequency (RF), or any other wired/wireless signaling mechanism.

The term “computerized” generally represents that any corresponding operations are conducted by hardware in combination with software and/or firmware.

Lastly, the terms “or” and “and/or” as used herein are to be interpreted as inclusive or meaning any one or any combination. Therefore, “A, B or C” or “A, B and/or C” mean “any of the following: A; B; C; A and B; A and C; B and C; A, B and C.” An exception to this definition will occur only when a combination of elements, functions, steps or acts are in some way inherently mutually exclusive.

As this invention is susceptible to embodiments of many different forms, it is intended that the present disclosure is to be considered as an example of the principles of the invention and not intended to limit the invention to the specific embodiments shown and described.

III. General Architecture

Referring to FIG. 1, an exemplary block diagram of a communication system 100 deploying a plurality of malicious content detection (MCD) systems 110₁-110_N (N>1, e.g. N=3) communicatively coupled to a management system 120 via a network 125 is shown. In general, management system 120 is adapted to manage MCD systems 110₁-110_N. For instance, management system 120 may be adapted to cause one or more clusters of objects, each of which comprise information representative of prior detected and classified objects, to be shared among some or all of the MCD systems 110₁-110_N for use in malware checks. Such sharing may be conducted automatically or manually uploaded by an administrator. Also, such sharing may be conducted freely among the MCD systems 110₁-110_N or subject to a subscription basis.

Herein, according to the embodiment illustrated in FIG. 1, a first MCD system 110₁ is an electronic device that is adapted to analyze information associated with network traffic routed over a communication network 130 between at least one server device 140 and at least one client device 150.

The communication network 130 may include a public computer network such as the Internet, in which case an optional firewall 155 (represented by dashed lines) may be interposed between communication network 130 and the client device 150. Alternatively, the communication network 130 may be a private computer network such as a wireless telecommunication network, wide area network, or local area network, or a combination of networks.

The first MCD system 110₁ is shown as being coupled with the communication network 130 (behind the firewall 155) via a network interface 160. The network interface 160 operates as a data capturing device (referred to as a “tap” or “network tap”) that is configured to receive network traffic propagating to/from the client device 150 and provide content from the network traffic to the first MCD system 110₁.

In general, the network interface 160 receives and duplicates the content that is received from and provided to client device 150 normally without an appreciable decline in performance by the server device 140, the client device 150, or the communication network 130. The network interface 160 may duplicate any portion of the content, for example, one or more files or objects that are part of a data flow or part of the payload contained within certain data packets, or the like.

In some embodiments, the network interface 160 may capture metadata from network traffic intended for the client device 150. This metadata may be used, at least in part, to deconstruct a corresponding file. For instance, the metadata may include keys that can be used to de-obfuscate a file or object.

It is contemplated that, for any embodiments where the first MCD system 110₁ is implemented as an dedicated appliance or a dedicated computer system, the network interface 160 may include an assembly integrated into the appliance or computer system that includes network ports, a network interface card and related logic (not shown) for connecting to the communication network 130 to non-disruptively “tap” network traffic propagating through firewall 155 and provide a copy of the network traffic to the dynamic analysis engine 190. In other embodiments, the network interface 160 can be integrated into an intermediary device in the communication path (e.g., firewall 155, router, switch or other network device) or can be a standalone component, such as an appropriate commercially available network tap. In virtual environments, a virtual tap (vTAP) can be used to duplicate files from virtual networks.

Referring still to FIG. 1, first MCD system 110₁ may include a scheduler 180, a storage device 185, a dynamic analysis engine 190, and a clustering and reporting module 195. In some embodiments, the network interface 160 may be contained within the first MCD system 110₁. Also, the dynamic analysis engine 190 and the clustering and reporting module 195 may be software modules executed by a processor that receives content and performs a dynamic scan analysis on objects within the content, which may involve accessing one or more non-transitory storage mediums operating as the storage device 185. In some embodiments, the dynamic analysis engine 190 may be one or more software modules, where such software modules are executed by a processor within the MCD system 110₁. The clustering and reporting module 195 may be one or more software modules executed by the same or a different processor, where these different processors are possibly located at geographically remote locations, located within the same processor package (e.g. different processor cores) and/or communicatively coupled for example via a network.

Herein, in one embodiment, the static analysis engine 175 may serve as a filter to permit subsequent malware analysis of one or more objects that may represent only on a portion of incoming content, which effectively conserves system resources and provides faster response time in determining the presence of malware within the analyzed content. As shown in FIG. 1, the static analysis engine 175 receives the copy of incoming content from the network interface 160 and applies heuristics to determine if any object(s) of the content are “suspicious”. The heuristics applied by the static analysis engine 175 may be based on data and/or rules stored in a database (not shown). Also, the static analysis engine 175 may examine the image of the captured content without executing or opening the captured content.

For example, the static analysis engine 175 may examine objects such as metadata or certain attributes of the captured content to determine whether a certain portion of the captured object matches (e.g., a high level of correlation with) a predetermined pattern of attributes that is associated with a malicious attack. According to one embodiment of the disclosure, the static analysis engine 175 flags objects from one or more data flows as suspicious after applying this heuristic analysis.

Thereafter, according to one embodiment of the invention, the static analysis engine 175 may be adapted to transmit at least an object of the suspicious content to the dynamic analysis engine 190. The portion of the object(s), such as some metadata for example, may identify attributes of the runtime environment in which the suspicious content should be processed and, on occasion, of the client device(s) 150 to which the suspicious content was being sent. Such metadata or attributes are used to identify a configuration of a virtual machine (VM) needed for subsequent malware analysis. In another embodiment of the disclosure, the dynamic analysis engine 190 may be adapted to receive one or more messages (e.g., data packets) from the static analysis engine 175 and analyze the message(s) to identify the software profile information associated with the needed VM.

For instance, as an illustrative example, the suspicious object(s) under test may include a portion of an email message that was generated, under control of Windows® 7 Operating System, using a Windows® Outlook 2010, version 1. Upon determining that the object includes suspicious content, such as an attachment for example, static analysis engine 175 provides software profile information to scheduler 180 to identify a particular configuration of VM needed to conduct dynamic analysis of the suspicious object. According to this illustrative example, the software profile information would include (1) Windows® 7 Operating System (OS); (2) Windows® Outlook 2010, version 1; and perhaps (3) an Adobe® reader if the attachment is a Portable Document Format (PDF) document.

The static analysis engine 175 supplies the software profile information to the scheduler 180, which determines whether any of the VM disk files within storage device 185 feature a software profile supporting the above-identified configuration of OS and one or more applications or a suitable alternative.

The dynamic analysis engine 190 is adapted to execute multiple VMs, to simulate the receipt and processing of different types of “suspicious” objects as well as different operating environments. Furthermore, the dynamic analysis engine 190 monitors and analyzes the activities and other behaviors of such objects during processing in the VM. The behaviors may include those expected and/or not expected during processing of that type of object. Unexpected behaviors can be considered anomalous behaviors. Examples of anomalous behaviors may include unusual network transmissions, opening certain ports to retrieve data, unusual changes in performance, and the like. This detection process is referred to as a dynamic malicious content detection.

The dynamic analysis engine 190 may flag the suspicious object as malware according to the observed behavior of the VM. In response to detecting anomalous behaviors, the dynamic analysis engine 190 may provide information to the cluster and reporting module 195 to conduct further analysis with objects previously analyzed and assigned to clusters, as described below.

Referring now to FIG. 2, an exemplary block diagram of logic associated with MCD system 110₁ is shown. MCD system 110₁ comprises one or more processors 200 that are coupled to communication interface logic 210 via a first transmission medium 220. Communication interface logic 210 enables communications with other MCD systems 110₂-110_N, management system 120 and/or cloud computing services 135 of FIG. 1. According to one embodiment of the disclosure, communication interface logic 210 may be implemented as a physical interface including one or more ports for wired connectors. Additionally, or in the alternative, communication interface logic 210 may be implemented with one or more radio units for supporting wireless communications with other electronic devices.

Processor(s) 200 is(are) further coupled to persistent storage 230 via transmission medium 225. According to one embodiment of the disclosure, persistent storage 230 may include static analysis engine 175, dynamic analysis engine 190, graphical user interface (GUI) logic 271, configuration logic 273, and clustering and reporting module 195, which comprises behavior analysis logic 231, sanitization logic 233, fuzzy hashing logic 235, comparison logic 237, and malware score logic 239. Of course, when implemented as hardware, engine 190 and logic 231, 233, 235, 237, 239, 271, and 273 would be implemented separately from persistent storage 230.

Turning now to FIG. 3, a method for classifying objects 300 will now be described. Each operation of the method 300 may be performed by one or more components of the MCD system 110₁. For example, the operation of method 300 may be performed by the dynamic analysis engine 190 in conjunction with the clustering and reporting module 195 of the MCD system 110₁. In other embodiments, the operations of method 300 may be performed in full or in part by other components of the communication system 100.

The method 300 may commence at operation 301 with receipt of a suspect object to be classified. The suspect object may be intercepted by the network interface 160 and passed to the MCD system 110₁ for analysis. In another embodiment, an anti-malware system running on the client device 150 may periodically and without direct provocation by the user intercept and transmit objects to the MCD system 110₁ for processing and analysis. This independent interception and analysis of objects allows the client device 150 to maintain an automatic examination of potential malware objects received without direct interaction by a user.

In another embodiment, a user of the client device 150 may submit a suspect object through a user interface. The interface may be generated by GUI logic 271 and served to the client device 150 by configuration logic 273 of the MCD system 110₁. In this fashion, the MCD system 110₁ may operate as a web-server to deliver data and a user interface to the client device 150.

FIG. 4A shows a web-interface 400 for submitting a suspect object to the MCD system 110₁ for analysis according to one embodiment. In this example interface 400, a user may direct a web browser running on the client device 150 to view the web-interface 400. The user may thereinafter enter the address/location of a suspect object into the web-interface 400 using the address input field 401 and the “BROWSE” button 403. The entered address indicates the location of the suspect object in storage on the client device 150 or on a remote device (e.g., stored on a website). After selection of a suspect object, the user may submit the suspect object for analysis by selecting the “SCAN” button 405 in the web-interface 400. The suspect object may be transmitted from the client device 150 such that it is received by the MCD 110₁ for processing as described above at operation 301.

In one embodiment, a suspect object may be any digital data structure. For example, a suspect object may be a file (e.g., PDF document), a component of a file, a component of a web page, an image, a series of captured network/web traffic that is capable of being replayed, etc. As described above, a user of the client device 150 may manually determine that an object is suspected to be malware or the client device 150 may automatically classify the object as potential/suspected malware and transmit the suspect object to the MCD system 110₁.

Referring back to FIG. 3, although described in relation to receiving a single suspect object, in other embodiments, the method 300 may be used in relation to multiple suspect objects received simultaneously or in rapid succession. For example, the method 300 may be used to analyze multiple suspect objects received from the client device 150 or other devices on the network 130. The suspect objects may be processed by the method 300 separately to determine whether each received suspect object is malware based on comparisons with previously generated clusters of objects using fuzzy hashes as described in greater detail below.

Following receipt of a suspect object, operation 303 determines behaviors of the suspect object using the behavior analysis logic 231. The determined behaviors characterize the suspect object such that a comparison can be performed with other previously classified objects in one or more object clusters as will be described in further detail below.

In one embodiment, the behaviors may be determined/detected at operation 303 after the suspect object has been detonated, namely processed (e.g. executed, opened or otherwise activated), by the dynamic analysis engine 190. For example, dynamic analysis engine 190 may detonate the suspect object such that operations associated with the suspect object are performed. For instance, in one embodiment the suspect object may be a PDF file. In this embodiment, the dynamic analysis engine 190 may detonate the PDF file by opening the file using an Adobe® Reader or other appropriate document reader.

In one embodiment, one or more virtual machines with various profiles that simulate the client device 150 may be used during detonation of the suspect object. These profiles may be software to be run by a virtual machine to process a suspect object. For example, the profiles may include an operating system and one or more suitable computer applications that are associated with the client device 150. For instance, an Adobe® Reader may be included in a virtual machine such that a suspect object, which is a PDF file, may be detonated by the virtual machine. Use of virtual machines ensures that detonation of the suspect object is controlled and will not result in infection of the client device 150 while still simulating the computing environment of the client device 150 to generate behavior data that describes the suspect object.

As noted above, detonation of the suspect object produces behavior data that describes the suspect object such that a comparison may later be performed with other objects. This behavior data may be detected and collected at operation 303 using the behavior analysis logic 231. The behavior data may include, for example, details regarding data generated by the suspect object during detonation, data accessed by the suspect object (both locally and from remote systems) during detonation, known exploits in the suspect object, etc.

In one embodiment, operation 305 may scrub the behavior data detected and collected at operation 303 to remove data that does not identify the suspect object. This scrubbing operation may be performed using the sanitization logic 233. In one embodiment, scrubbing the behavior data includes removing a subset of process identifiers of processes called by the suspect object during detonation, values written to, deleted from, or modified to a registry by the suspect object during detonation such that only the path of these operations is retained, and names of files generated, modified, and/or deleted by the suspect object during detonation such that only a path in an associated file system is retained. This removed/scrubbed data may be discarded at operation 305 as it does not identify the suspect object in relation to other objects and may be considered superfluous.

After the behavior data has been scrubbed at operation 305 to generate scrubbed behavior data, the method 300 may perform two analyses: 1) an analysis to associate the suspect object with a cluster of previously stored/analyzed objects and 2) an analysis to generate a malware score, which describes the probability that the suspect object is malware. The analyses may be run concurrently or asynchronously. In one embodiment, the results of the first analysis (i.e., cluster association) may be used to modify the malware score generated by the second analysis. Each of these analyses will be described in greater detail below.

Beginning with the first analysis of the suspect object, at operation 307 a fuzzy hash for the suspect object may be generated based on the scrubbed behavior data using the fuzzy hashing logic 235. A fuzzy hash allows the comparison of objects to determine similarity of the objects instead of necessarily a direct match. In comparison, traditional hashing techniques only allow a comparison of objects to determine an exact match. By allowing the determination of “similar” objects, fuzzy hashes afford a greater leniency in classification and categorization of objects that might be slightly different but otherwise share important characteristics. Through the utilization of a fuzzy hash, similar objects may be determined through a comparison of hash values within the fuzzy hash as will be described in greater detail below.

In one embodiment, a fuzzy hash is constructed by running a hashing algorithm over blocks of the scrubbed behavior data for an object. In one embodiment, an MD5 hash may be performed on successive blocks of scrubbed behavior data to produce a plurality or a stream of hash values. For example, the scrubbed behavior data may be separated into N equal sized blocks, where N is greater than or equal to two (e.g., 1024 byte blocks). A hash value is produced for each of the N blocks to generate exactly N hash values. In one embodiment, the scrubbed behavior data may be separated into blocks corresponding to segments of data that represent discrete behaviors detected at operation 303. Accordingly, in this embodiment, each block represents a single detected behavior associated with the suspect object.

Although described in relation to use of an MD5 hash for generation of the fuzzy hash, in other embodiments other hashing techniques/methods may be used. For example, in other embodiments a SHA, SWIFFT, and/or HAVAL hash may be used to generate the fuzzy hash for the suspect object at operation 307.

Following the generation of a fuzzy hash for the suspect object at operation 307, operation 309 may compare the fuzzy hash for the suspect object with one or more fuzzy hashes of other previously stored/classified objects associated with clusters. This comparison may be performed using the comparison logic 237 of the MCD system 110₁. In one embodiment, the previously stored clusters of objects are stored locally on the MCD system 110₁ in the storage device 185 or a separate data store (e.g. part of persistent storage 230 of FIG. 2). In other embodiments, the previously stored clusters of objects may be stored in cloud computing services 135 or the management system 120. In these embodiments, the management system 120 may distribute clusters of objects to MCD systems 110₁-110₃ as needed or the MCD systems 110₁-110₃ may directly access the clusters of objects over the network 125 for analysis of other objects received by these MCD systems 110₁-110₃.

As shown in FIG. 5, each previously stored object 501 may be associated with one or more behaviors. Similar to the behaviors of the suspect object, the behaviors of the previously stored objects 501 characterize the dynamic actions, operations, and activities of the objects 501 during detonation. A fuzzy hash may be associated with each object 501 based on these behaviors in a similar fashion as described above in relation to operation 307. Each of the previously stored objects 501 may be associated with a cluster 503 based on a similarity of fuzzy hashes for each respective object 501. For example, as shown in FIG. 5, objects 501₁ and 501₂ are associated with the cluster 503₁. This association indicates that the fuzzy hashes of objects 501₁ and 501₂ are “similar” and may be considered in the same family. In one embodiment, similarity may be described in terms of the number of matching hash values between the respective fuzzy hashes of objects 501₁ and 501₂. For example, the fuzzy hashes of two objects may be compared to determine a similarity measure. The similarity measure may describe the percentage or number of matching hash values between the two fuzzy hashes. In one embodiment, a similarity measure above a predefined similarity threshold indicates that the objects are similar. Since the objects 501₁ and 501₂ are in the same cluster 503₁, the comparison of fuzzy hashes for objects 501₁ and 501₂ would yield a similarity measure above the predefined similarity threshold. In contrast, the comparison of fuzzy hashes for objects 501₁ and 501₃ would yield a similarity measure below the predefined similarity threshold since these objects 501₁ and 501₃ are associated with different clusters (e.g., clusters 503₁ and 503₂ respectively).

Referring back to FIG. 3 and returning to operation 309, the fuzzy hash of the suspect object is compared with one or more fuzzy hashes of previously stored objects associated with clusters. Using the example set of objects 601 in FIG. 6, the fuzzy hash of the suspect object may be compared against the fuzzy hashes of one or more of the objects 601₁-601₁₅. In one embodiment, operation 309 compares the fuzzy hash of the suspect object with the fuzzy hash of at least one object 601 in each cluster 603. Each comparison yields a separate similarity measure that describes the similarity of the suspect object and each respective comparison object 601.

In one embodiment, operation 309 may utilize statistical and machine learning to determine whether the suspect object is similar to an object in a cluster. Machine learning refers to a process or system that can learn from data, i.e., be trained to distinguish between “good” and “bad”, or in this case, between similar objects and non-similar objects. The core of machine learning deals with representation and generalization, that is, representation of data objects (e.g., the anomalies and other analytical results, which can be collectively represented by features/behaviors of the objects), and functions performed on those objects (e.g., weighting and probability formulas). Generalization is the property that the process or system uses to apply what it learns on a learning set of known (or “labeled”) data objects to unknown (or “unlabeled”) examples. To do this, the process or system must extract learning from the labeled set that allows it to make useful predictions in new and unlabeled cases.

For machine learning, the MCD system 110₁ may operate in a training mode and in an operational mode. In a training mode, the MCD system 110₁ employs threat heuristics training logic to subject known samples (e.g., labeled samples) of similar objects and known samples of non-similar objects to calibrate threat heuristics logic for probability scoring and/or decision making of objects. To accomplish this, the threat heuristics training logic may submit similar and non-similar objects to analyzers. In some embodiments, the threat heuristics training logic may employ a special forensics system. In alternative embodiments, the threat heuristics training logic may test the similar and non-similar objects each time it processes a different suspect object, or it may store the results of prior tests for use for future processing of objects. The threat heuristics training logic may assign a probability score (e.g., a similarity measure) to each of the possible patterns resulting from testing the similar and non-similar objects. These probability scores and classification labels are indicative of whether a set of objects are similar. In one embodiment, the machine learning routines and operations described above may be performed by the learning module 187 shown in FIG. 1 based on inputs from the storage device 185 and/or the clustering and reporting module 195.

Referring back again to FIGS. 2-3, at operation 311, the set of similarity measures generated at operation 309 may be compared against the predefined similarity threshold to determine whether the suspect object is “similar” to a previously stored object in a preexisting cluster. This comparison may be performed by the comparison logic 237 and reveals whether the suspect object is within the same family as objects within a cluster. As noted above, if a similarity measure is above the predefined similarity threshold, the suspect object is “similar” to the corresponding object. However, if the similarity measure is below the predefined similarity threshold, the suspect object is not “similar” to the corresponding object. In one embodiment, the predefined similarity threshold may be set by an analyst, network administrator, and/or subscriber.

Upon determining that a similarity measure is above the predefined similarity threshold, the method moves to operation 313. At operation 313, the suspect object is associated with the cluster of the object with which the generated similarity measure exceeded the predefined similarity threshold. For example, as shown in FIG. 7A, the fuzzy hash of the suspect object may be added to cluster 603₂. In this example, the similarity measure between the suspect object and one or more of the objects 601₅-601₈ is above the predefined similarity threshold.

In one embodiment, association with a cluster may be used to further describe the suspect object. For example, association with a cluster may be used to 1) determine a malware family name for the suspect object and/or 2) determine whether the suspect object is malware, non-malware, or has an unknown status.

As shown in FIG. 7A, the objects 601₁-601₄ in the cluster 603₁ were determined to be non-malware (indicated by the lack of shading for these objects 601₁-601₄). This classification determination may be based on a previous dynamic or static analysis of the objects 601₁-601₄. In this case, if the suspect object had been associated with the cluster 603₁, the suspect object would be classified as non-malware.

In comparison, the objects 601₅-601₈ in the cluster 603₂ were determined to be malware (indicated by shading of these objects 601₅-601₈) and associated with the malware family name “MalBot”. Again, this classification determination may be based on a previous dynamic or static analysis of the objects 601₅-601₈ using both comparisons with locally stored objects and objects stored remotely. Since the suspect object has been associated with the cluster 603₂ in the example provided above, the suspect object is classified malware and associated with the malware family name “MalBot”.

In some instances, a status of a set of objects in a cluster may not yet be known. For example, in the cluster 603₄ shown in FIG. 7A the status of these objects 601₁₂-601₁₆ cannot yet be determined as malware or non-malware (indicated by dashed border for these objects 601₁₂-601₁₆). Accordingly, if the suspect object had been associated with the cluster 603₄, the suspect object would be classified with an unknown status. Grouping objects with unknown status may later be useful when a classification and/or malware family name may be assigned to these objects.

In some embodiments, association of an object with a cluster may only be informative and not provide classification information. For example, the cluster 603₃ may include several objects 601₉-601₁₁ that have been classified as malware and associated with the malware family name “DataStealer”. However, association with cluster 603₃ may only yield an association with a malware family name associated with the cluster 603₃ (e.g., “DataStealer”) instead of also a classification for the newly added object. This failure to yield classification information for new objects may be based on a number of false positive malware classifications associated with the cluster 603₃ or another threshold that indicates an unreliable classification.

In one embodiment, operations 309 and 311 may be first performed in relation to clusters of objects stored locally on the MCD system 110₁ (i.e., in the storage device 185). Following a failure to locate a locally stored cluster with a “similar” object to the suspect object, the operations 309 and 311 may be performed for clusters of objects stored on other devices. For example, the operations 309 and 311 may be performed on clusters of objects stored on a cloud server located in the cloud computing services 135 in response to a failure to locate a local cluster with a “similar” object.

Returning to operation 311 of FIG. 3, upon determining that similarity measures generated for the suspect object at operation 309 are not above the predefined similarity threshold for any local or remote clusters of objects, the method 300 moves to operation 315 to create a new cluster for the suspect object. For example, FIG. 7B, shows the suspect object added to new cluster 603₅. In this example, the suspect object is not similar to any of the objects 601₁-601₁₆ based on compared fuzzy hashes and accordingly is not part of these families of objects. Instead, the suspect object is the first member of a new family defined by the cluster 603₅.

Following generation of a new cluster for the suspect object at operation 315, operation 317 may transmit the new cluster to the MCD systems 110₂ and 110₃ and/or the management system 120. In one embodiment, the management system 120 may receive the new cluster from the MCD system 110₁ and propagate this new cluster to the MCD systems 110₂ and 110₃ using the network 125. The MCD systems 110₂ and 110₃ may utilize this new cluster for future analysis of other objects intercepted or otherwise received from the client device 150 or other devices on the network 130.

As described above, objects intercepted or otherwise received from the client device 150 may be compared using fuzzy hashes to determine similarity. Upon determination of similarity, the received/suspect object may be associated with a corresponding cluster and inherit attributes of the cluster. These attributes may include 1) classification as malware, non-malware, or an unknown status and/or 2) a malware family name. By utilizing fuzzy hash comparisons with previously stored and classified objects, the method 300 provides an efficient technique for classifying newly received objects based on familial similarities.

In one embodiment, the results of the method 300 may be transmitted from the clustering and reporting module 195 to the dynamic analysis engine 190. In this embodiment, the results of the method 300 may be used to supplement the analysis results produced by the dynamic analysis engine 190 to increase the accuracy in identifying suspicious objects as malware.

As noted above, the method 300 may conduct a separate analysis following operation 305 to generate a preliminary malware score, which describes the probability that the suspect object is malware. For example, the preliminary malware score may fall between 0.0 and 1.0. In one embodiment, operation 319 compares the scrubbed behavior data of the suspect object with known malware behaviors using the malware score logic 239 shown in FIG. 2. These known malware behaviors may be cultivated after dynamic analysis of known malware objects by the MCD 110₁, another device on the network 125 (e.g., the MCDs 110₂ and 110₃ or the management system 120), and/or a remote device (e.g., device located within the cloud computing services). In one embodiment, the known malware behaviors are stored in the storage device 185 and describe unexpected, anomalous, and/or malicious actions that are characteristic of malware. Examples of anomalous behaviors may include unusual network transmissions, opening certain ports to retrieve data, unusual changes in performance, and the like.

The comparison at operation 319 yields a preliminary malware score based on the number of similarities between the scrubbed behavior data and the known malware behavior. For example, when multiple behaviors described in the scrubbed behavior data match behaviors in the known malware behaviors, operation 319 may yield a high preliminary malware score (e.g., 0.9), which indicates a high probability the suspect object is malware. In contrast, when few behaviors described in the scrubbed behavior data match behaviors in the known malware behaviors, operation 319 may yield a low preliminary malware score (e.g., 0.1), which indicates a low probability the suspect object is malware. In one embodiment, this comparison at operation 319 may be performed using machine learning and statistical analysis similar to that described above in relation to operation 309.

In one embodiment, the preliminary malware score may be used at operation 321 to generate a final malware score based on the suspect object's association with a cluster at operations 313 or 315. For example, when the suspect object is associated with a cluster that classifies the suspect object as malware, the preliminary malware score from operation 319 may be increased to generate a final malware score that is greater that the preliminary malware score from operation 319. This increase indicates a higher probability that the suspect object is malware than originally computed at operation 319. Conversely, when the suspect object is associated with a cluster that classifies the suspect object as non-malware or with an unknown status, the preliminary malware score from operation 319 may be decreased to generate the final malware score. This decrease indicates a lower probability that the suspect object is malware than originally computed at operation 319. By generating a final malware score that reflects the probability that a suspect object is malware based on both a comparison with known malware behaviors and clusters of classified objects, operation 321 creates a more robust determination of the likelihood that the suspect object is malware.

At operation 323, the final malware score generated at operation 321 may be transmitted along with the classification and naming information assigned to the suspect object at operations 313 or 315 to a user of the client device, a subscriber of a malware detection service, a network administrator, or another entity. The transmission may be made using an email message, a popup message, or any other message transmission technique. For example, the user interface 400 may be updated to reflect the classification of the suspect object as shown in FIG. 4B.

As described above, the method for classifying objects 300 may utilize fuzzy hash techniques to group “similar” objects in clusters for future analysis. This similarity matching allows greater flexibility in analyzing potential malware objects, which may share multiple characteristics and behaviors but are also slightly different from previously classified objects. These clusters of objects may be continually updated and shared between the MCD systems 110₁-110_N as new objects are processed by the method 300 such that a robust set of object clusters are maintained for future detection and remediation of families of malware threats.

Claims

1. A computerized method for classifying objects in a system configured to detect malicious content within one or more objects analyzed by the system, comprising:

detecting, by the system, behaviors of an object for classification after processing of the received object has started;

collecting data associated with the detected behaviors;

generating a fuzzy hash for the received object based on the data associated with the detected behaviors, the generating of the fuzzy hash includes:

(i) removing a portion of the data associated with the detected behaviors to produce a remaining portion of the data associated with the detected behaviors, and

(ii) performing a hash operation on the remaining portion of the data associated with the detected behaviors;

comparing the fuzzy hash for the received object with a fuzzy hash of an object in a preexisting cluster to generate a similarity measure;

associating the received object with the preexisting cluster in response to determining that the similarity measure is above a predefined threshold value; and

reporting, by the system via a communications interface, whether the received object is associated with the preexisting cluster.

2. The computerized method of claim 1, further comprising:

creating a new cluster for the received object in response to determining that the similarity measure is below the predefined threshold value.

3. The computerized method of claim 1, wherein the received object is at least one of a file, a uniform resource locator, a web object, a capture of network traffic for a user over time, and an email message.

4. The computerized method of claim 1, wherein the changes performed by the received object include (1) network calls, (2) modifications to a registry, (3) modifications to a file system, or (4) an application program interface call.

5. The computerized method of claim 1, further comprising:

generating a preliminary malware score for the received object based on a comparison of the detected behaviors with known malware behaviors, wherein the preliminary malware score indicates the probability the received object is malware; and

generating a final malware score for the received object based on the cluster the received object is associated, wherein the final malware score is greater than the preliminary malware score when the received object is associated with a cluster of objects classified as malware and the final malware score is less than the preliminary malware score when the received object is associated with a cluster of objects classified as non-malware.

6. The computerized method of claim 1, wherein the removing of the portion of the data associated the detected behaviors comprises removing data that does not identify the received object.

7. The computerized method of claim 6, wherein the removing of the portion of the data associated the detected behaviors includes at least one of: (i) a process identifier of a process called by the received object, or (ii) a value written to a registry by the received object, or (iii) names of objects generated, modified, or deleted by the received object.

8. The computerized method of claim 2, further comprising:

transmitting, by the system, the new cluster or the preexisting cluster with the newly associated received object to another system configured to detect malicious content within one or more objects.

9. The computerized method of claim 1, further comprising:

classifying the received object as malware, non-malware, or with an unknown status to match a classification of the preexisting cluster, when the received object is assigned to the preexisting cluster.

10. The computerized method of claim 1, further comprising:

assigning a malware family name to the received object to match a malware family name of the preexisting cluster, when the received object is assigned to the preexisting cluster.

11. The computerized method of claim 1, wherein the removing of the data associated with the detected behaviors includes removing values written to a registry or modified registry values.

12. The method of claim 1, wherein the behaviors of the object including the system processing a suspected malware object by detonating, executing or opening the suspected malware object to cause the suspected malware object to perform the behaviors.

13. The method of claim 1, wherein the pre-existing cluster is retrieved from cloud computing services.

14. The method of claim 1, wherein the received object is provided from a client device remotely located from the system.

15. A system comprising:

one or more hardware processors;

a memory including one or more software modules that, when executed by the one or more hardware processors:

detect behaviors of a received object for classification after processing of the received object has started;

collect data associated with the detected behaviors;

generate a fuzzy hash for the received object based on the data associated with the detected behaviors, the fuzzy hash being generated by at least:

(i) removing a portion of the data associated with the detected behaviors to produce a remaining portion of the data associated with the detected behaviors, and

(ii) performing a hash operation on the removed portion of the data associated with the detected behaviors;

compare the fuzzy hash for the received object with a fuzzy hash of an object in a preexisting cluster to generate a similarity measure;

associate the received object with the preexisting cluster in response to determining that the similarity measure is above a predefined threshold value; and

report whether the received object is associated with the preexisting cluster.

16. The system of claim 15, wherein the one or more software modules, when executed by the one or more hardware processors, are further configured to:

create a new cluster for the received object in response to determining that the similarity measure is below the predefined threshold value.

17. The system of claim 15, wherein the removing the data associated with the detected behaviors includes removing data associated with the detected behaviors that does not identify the received object, wherein the fuzzy hash is generated using the detected behaviors.

18. The system of claim 15, wherein the one or more software modules, when executed by the one or more hardware processors, are further configured to:

classify the received object as malware, non-malware, or with an unknown status to match a classification of the preexisting cluster, when the received object is assigned to the preexisting cluster.

19. The system of claim 15, wherein the one or more software modules, when executed by the one or more hardware processors, are further configured to:

assign a malware family name to the received object to match a malware family name of the preexisting cluster, when the received object is assigned to the preexisting cluster.

20. The system of claim 15 being in communication with a client device remotely located from the system to receive the received object.